bred vector and enso predictability in a hybrid coupled
TRANSCRIPT
Bred Vector and ENSO Predictability in a Hybrid Coupled Model during thePeriod 1881–2000
YOUMIN TANG AND ZIWANG DENG
Environmental Science and Engineering, University of Northern British Columbia,
Prince George, British Columbia, Canada
(Manuscript received 27 October 2009, in final form 2 July 2010)
ABSTRACT
In this study, a breeding analysis was conducted for a hybrid coupled El Nino–Southern Oscillation (ENSO)
model that assimilated a historic dataset of sea surface temperature (SST) for the 120 yr between 1881 and
2000. Meanwhile, retrospective ENSO forecasts were performed for the same period. For a given initial state,
15 bred vectors (BVs) of both SST and upper-ocean heat content (HC) were derived. It was found that the
average structure of the 15 BVs was insensitive to the initial states and independent of season and ENSO
phase. The average structure of the BVs shared many features already seen in both the final patterns of
leading singular vectors and the ENSO BVs of other models. However, individual BV patterns were quite
different from case to case. The BV rate (the average cumulative growth rate of BVs) varied seasonally, and
the maximum value appeared at the time when the model ran through the boreal spring and summer. It was
also sensitive to the strength of the ENSO signal (i.e., the stronger ENSO signal, the smaller the BV rate).
Furthermore, ENSO predictability was explored using BV analysis. Emphasis was placed on the re-
lationship between BVs, which are able to characterize potential predictability without requiring observa-
tions, and actual prediction skills, which make use of real observations. The results showed that the relative
entropy, defined using breeding vectors, was a good measure of potential predictability. Large relative en-
tropy often leads to a good prediction skill; however, when the relative entropy was small, the prediction skill
seemed much more variable. At decadal/interdecadal scales, the variations in prediction skills correlated with
relative entropy.
1. Introduction
Intensive research has been focused on studying pre-
dictability, including potential predictability and actual
prediction skill, by examining the perturbation growth.
Currently, there are two methods, the singular vector
(SV) and the bred vector (BV), which are widely used to
analyze the optimal perturbation growth. The earliest
work on SVs can be found in Lorenz’s (1965) paper,
which introduced SVs into meteorology for atmospheric
predictability studies. However, SV analysis was not used
to investigate ENSO predictability until the late 1990s.
Recently, a number of studies on ENSO predictability
have used SVs (e.g., Xue et al. 1997a,b; Chen et al. 1997;
Thompson 1998; Moore and Kleeman 1996, 1997a,b; Fan
et al. 2000; Tang et al. 2006; Zhou et al. 2007).
The breeding method was originally developed to
‘‘bred’’ growing modes related to synoptic baroclinic
instability for short-to-medium-range atmospheric en-
semble forecasting by taking advantage of the fact that
small-amplitude, but faster-growing, convective insta-
bilities saturate earlier. It was first proposed in the 1990s
by Toth and Kalnay (1993, 1997). Breeding vectors are,
by construction, closely related to Lynapunov vectors
(LVs). While BVs have been widely used in mesoscale
and synoptic-scale predictions (e.g., Corazza et al. 2003;
Kalnay 2003; Young and Read 2008), they have rarely
been applied to climate predictions. Cai et al. (2003,
hereafter CKT03) first applied the breeding method
to the Zebiak and Cane (ZC) (1987) ENSO model, and
investigated features of the ENSO BV modes and the
potential of BVs for use in data assimilation and ensem-
ble prediction. Yang et al. (2006, hereafter Y06) explored
the possibility of separating ENSO-related coupled vari-
ability from weather modes and other shorter-time-scale
oceanic instabilities using the breeding technique with
a fully coupled general circulation model (GCM). Some
Corresponding author address: Youmin Tang, Environmental
Science and Engineering, University of Northern British Colum-
bia, 3333 University Way, Prince George, BC V2N 4Z9, Canada.
E-mail: [email protected]
298 J O U R N A L O F C L I M A T E VOLUME 24
DOI: 10.1175/2010JCLI3491.1
� 2011 American Meteorological Society
important properties of ENSO BVs were discovered in
CKT03 and Y06, including the relationship of BVs to the
ENSO background. They also found that a BV pattern is
very similar to the final pattern of the singular vector of
the same coupled model.
Several issues of ENSO BVs still warrant further ex-
ploration. First, both CKT03 and Y06 used model in-
tegration as the reference trajectory (i.e., the initial
state) from which a BV is calculated. They performed
BV analysis continuously during the period of interest
from the beginning to the end. It is also interesting to
explore ENSO BVs by using the actual ENSO evolu-
tions as the reference trajectories, which reveals the
properties and features of ENSO BVs based on realistic
ENSO backgrounds. In particular, the realistic refer-
ence trajectory allows for the exploration of the re-
lationship between BVs, which characterize potential
predictability without requiring observations, and actual
prediction skills, which make use of real observations.
Second, it is well known that ENSO predictability has
a striking decadal/interdecadal variation (e.g., Chen
et al. 2004; Deng and Tang 2009; Tang et al. 2008a). One
might be able to shed some light on the mechanism of
this decadal/interdecadal variation in ENSO predict-
ability by examining temporal variations in potential
predictability, measured by the BV rate or other BV de-
rivatives. In addition, CKT03 and Y06 used, respectively,
an intermediate coupled model and a fully coupled GCM
for ENSO BV analysis. Applying the breeding method
into a hybrid coupled model facilitates ENSO BV analysis
for a full spectrum of models.
Considering the use of the breeding method in a hy-
brid coupled model for the above goals, the period of
BV analysis with a realistic ENSO background should
be long enough to cover sufficient ENSO cycles, allow-
ing statistically robust conclusions to be drawn. In ad-
dition, a long-term retrospective ENSO prediction is
required, in order to examine the relationship between
BVs and actual prediction skill. The relationship between
potential predictability and actual prediction skills has a
practical significance, and may offer a means of estimat-
ing the confidence that we can place in predictions of the
future using the same climate model.
This study pursues the aforementioned objectives. We
recently reconstructed the surface wind stress of the
tropical Pacific back to 1881 using statistical techniques
and using a historic SST and sea level pressure datasets
(Deng and Tang 2009). The reconstructed wind stress
was applied to a hybrid coupled model to perform ret-
rospective ENSO predictions from 1881 to 2000 (Deng
and Tang 2009; Tang et al. 2008a). We also completed
the assimilation of a long-term historic SST dataset into
an oceanic general circulation model (OGCM) that
generated skillful retrospective predictions (Deng et al.
2008). These steps made it possible to implement BVs
with realistic ENSO backgrounds over a long period
and to explore the relationship between BVs and pre-
dictability. This paper is structured as follows: the model
and breeding method are introduced in section 2; the re-
sults on ENSO breeding modes and optimal error growth
rates are presented in section 3; in section 4, the associa-
tion between the features of BVs and ENSO predictability
is explored—especially the measurement of potential
predictability based on BVs; finally, the discussion and
conclusions can be found in section 5.
2. Model and breeding method
a. Model
A hybrid coupled model (HCM) that was composed
of an OGCM coupled to a linear statistical atmospheric
model was used for this study. The OGCM was the latest
version of the Nucleus for European Modeling of the
Ocean (NEMO), identical to the OGCM used in Tang
et al. (2008a) and Deng and Tang (2009). (Details of the
ocean model are available at http://www.lodyc.jussieu.
fr/NEMO/.) The domain of the model used here is
configured for the tropical Pacific Ocean from 308N to
308S and 1228E to 708W, with horizontal resolution 2.08
in the zonal direction and 0.58 within 58 of the equator,
smoothly increasing to 2.08 at 308N and 308S in the me-
ridional direction. There were 31 vertical levels with 17
concentrated in the top 250 m of the ocean. The time step
of integration was 1.5 h and all boundaries were closed,
with no-slip conditions. The statistical atmospheric model
in the HCM was a linear model that predicted the con-
temporaneous surface wind stress anomalies from SST
anomalies (SSTA), which are constructed by the singu-
lar vector decomposition (SVD) method with a cross-
validation scheme. During the initialization of the hybrid
coupled model, the OGCM was forced by the sum of the
associated wind anomalies computed from the atmo-
spheric model and the observed monthly mean climato-
logical wind stress. Full details of the model are given in
Deng et al. (2008) and Tang et al. (2005, 2008a).
To perform long-term hindcasts with the HCM, past
wind stress data are required to initialize forecasts. Us-
ing SST as a predictor, as well as SVD techniques, we
reconstructed a wind stress dataset back to 1881 (Deng
and Tang 2009). The SST data came from the monthly
extended global SST (ERSST) dataset from 1881 to 2000
reconstructed by Smith and Reynolds (2004), and had
28 latitude by 28 longitude resolution. The surface wind
data came from the monthly National Centers for En-
vironmental Prediction–National Center for Atmospheric
1 JANUARY 2011 T A N G A N D D E N G 299
Research (NCEP–NCAR) reanalysis dataset from 1948
to 2000. The data domains for both SST and wind were
configured to the tropical Pacific Ocean. The recon-
structed wind data has been used to study ENSO pre-
dictability (e.g., Tang et al. 2008a; Deng and Tang 2009).
The reconstructed winds were assimilated with SST and
the HCM to produce the initial conditions of predictions
using an ensemble Kalman filter (EnKF; Deng et al. 2008).
The retrospective forecasts for the period from 1881 to
2000 were analyzed in detail in Tang et al. (2008a).
b. Breeding analysis
In this study, we derived a BV using the two-sided self-
breeding algorithm. The nonlinear model is denoted by
M; X indicates the model states that are perturbed for
the BV. The BV at step k 1 1 can be obtained from the
below equations, as schematically described by Fig. 1:
BVk11
5 DX�k11, (1)
DX�k11 51
2
ðtk1t
tk
M(Xk
1 DX1k ) dt
"
�ðt
k1t
tk
M(Xk� DX1
k ) dt
#, (2)
DX1k11 5
DX�k11
lk11
, and (3)
lk11
5B�k11
B1k
, (4)
where k 5 0, 1, . . . , N. The number N is the last rescaling
step. The term lk11 is the rescaling factor at step k 1 1,
which is defined by the amplitude of the perturbation
growth of the upper-ocean 250-m heat content (HC)
over the region of Nino-3.4 (58N–58S, 1708–1208W). The
upper HC was chosen here because of its critical im-
portance in ENSO simulation and prediction (e.g., Tang
2002). However CKT03 found that a BV is, in fact, rel-
atively insensitive to the definition of the rescaling norm.
The variables B�k11 and Bk1 in (4) are the spatially
averaged amplitude of DHC�k11 and DHCk1 of the upper-
ocean 250-m HC over the domain of Nino-3.4, re-
spectively. Note that the HC is only a model diagnostic
variable, thus DHC�k11 is obtained using the HC defini-
tion and (2). In this model, the upper 250 m is covered
by the first 17 layers, and the HC is defined as below
HC 5
�17
i51T
ih
i
�17
i51h
i
, (5)
where Ti and hi are the temperature and thickness, re-
spectively, of the ith model layer obtained by the model
integration [i.e., the first or second item of rhs of (2)] and
the model initial configuration in vertical, respectively.
If we denote DHC by the one-dimensional vector Dhc(i),
i 5 1, 2, . . . , N1, where N1 is the number of model grids
over Nino-3.4 region, we have
B�k11 5
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
N1�N1
i51Dhc2(i)�k11
vuutand (6)
B1k 5
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
N1�N1
i51Dhc2(i)1
k
vuut. (7)
Initially, k 5 0, DX01is the initial random perturbation
x, and B01 is the amplitude of initial perturbation x. The
term t is called the rescaling interval. One round of
process from k to k 1 1 is called one breeding cycle, or
a rescaling step. A steady BV is obtained after a fast
growth of dominant instabilities is saturated (e.g., the
BVs after K 1 L in Fig. 1a).
There are two important parameters in deriving a BV:
the initial perturbation amplitude and t. The initial per-
turbation magnitude for this study was randomly drawn
from a normal distribution with a mean of zero and a
variance of a specified value that was 10% of the variance
of the model heat content anomalies of the upper ocean
250-m HC, that is,
B10 ; N[0, 0.1s(HC)], (8)
where s(HC) is the spatially averaged variance of HC
over the region of Nino-3.4, which was obtained from
the model control run from 1961 to 2000. Thus, the ini-
tial perturbation field x is determined by
x 5 B10 V. (9)
The matrix V is the normalized pseudo-Gaussian
noise field with spatial coherence, given by Evensen
(2003), which has been widely used in ensemble Kalman
filter. Several sensitivity experiments show that, when
using this setting for the initial perturbation, the per-
turbed solution quickly diverges from the control solution
after one rescaling step and meanwhile, the breeding
vectors can converge in four–six rescaling steps, as shown
in Fig. 2.
The rescaling interval is a little arbitrary. Pena and
Kalnay (2004) argued that this period should be longer
than two weeks for slowly varying coupled instability.
CKT03 used three months to obtain bred vectors for the
ZC model whereas Y06 used one month for a coupled
300 J O U R N A L O F C L I M A T E VOLUME 24
global general circulation model. In this study, we choose
a rescaling period of one month, which leads to the sat-
uration of fast growing instabilities after four–six rescal-
ing steps, as discussed in the next section.
To provide a sufficient sample of the dynamically
important uncertainties of the initial state, we used 15
pairs of random perturbations for each initial condition
(i.e., obtaining 15 BVs). The perturbed variables in-
cluded the sea temperature, horizontal currents, and
vertical currents of the top 250 m (17 model levels).1
The BVs were computed at 3-month intervals from
January 1881 to October 2000 (i.e., January 1881, 1 April
1881, . . . October 2000). Thus there are a total of 15 3
480 BVs for each variable of interest during the entire
period, providing sufficient sampling to conduct robust
statistical analyses. The corresponding time for each BV
was the initial time from which the random initial per-
turbation was added to the model state (see Fig. 1b).
It should be mentioned that the above BV approach is
different from that used in CKT03 and Y06 because they
analyzed BV cycles continuously until the end of the
entire period, as demonstrated in Fig. 1a. The essential
difference between our approach and theirs is the initial
perturbation and initial condition used to generate the
BV. In their approach, the rescaling is done continu-
ously until the end. The initial perturbation for the next
BV was always the previous BV, and was imposed onto
unperturbed forecasts, ensuring the fast growth of in-
stabilities to be saturated in one rescaling step after
a period of initial training (e.g., beyond K 1 L in Fig.
1a). It is similar to the method of Toth and Kalnay
(1993) where the previous BV was centered around the
evolving analysis state for the next BV. A common
feature of CKT03, Y06, and Toth and Kalnay (1993) is
that the analysis state (or unperturbed forecast) and BV
analysis are for the same model. However, the analysis
state in this work is from an ocean data assimilation
system that is different from the coupled model used for
BV analysis; thus, inconsistencies exist between mod-
eling system for the BV cycle and the modeling system
for oceanic analyses, leading to an initial shock in per-
forming BV of each initial condition. Thus, our ap-
proach always takes a complete BV analysis from a
random initial perturbation to the saturation of the fast
FIG. 1. A schematic of the two-sided self-breeding BV algorithm. (a) The steady BVs were
obtained only after a few rescaling steps (i.e., K 1 L) and the unperturbed model forecast was
always used as a reference trajectory for each BV. (b) Each BV was derived from a realistic
initial condition and obtained by six rescaling steps (see text). The interval between two re-
scaling steps (e.g., k to k 1 1) was 1 month. In this study, the BV was analyzed every 3 months
from January 1881 to October 2000 using (b).
1 Considering the computational expense in preparing initial
perturbations, we did not perturb all model variables. The choice of
temperature and currents as the variables that were perturbed was
due primarily to their significant importance in ENSO formulation
and development although it was a somewhat arbitrary. The top
250 m cover the thermocline in the equator, which dominates
ENSO variability.
1 JANUARY 2011 T A N G A N D D E N G 301
growth of instabilities for each analysis state (Fig. 1b).
Sensitivity experiments indicate that such a complete
BV analysis usually takes four to six rescaling steps for
the HCM that has a noise-free atmospheric component.
Such an approach also allows us to analyze the BV of
each initial condition in a completely parallel manner. In
concept, our approaches is the same as that of Toth and
Kalnay (1993), including its definition of a BV, although
it is more expensive. We used a realistic analysis state,
rather than an unperturbed forecast, as the initial condi-
tion of each BV because we wanted to explore pertur-
bation growth under a realistic ENSO background, in
addition to the relationship between BV s and prediction
skills. In fact, our approach is very similar to the singular
vector analysis for each initial condition that was per-
formed by Lorenz (1965) using a linear tangent model.
To measure the perturbation growth from the initial
time until the saturation of fast-growing instabilities, a
cumulative growth factor for breeding vector m after step
k was defined, as described in Young and Read (2008):
Gmk 5
lm1 if k 5 1;
lmk Gm
k�1 if k . 1.
�
The term Gkm is lk multiplied by the previous cumu-
lative growth rate. A cumulative growth factor may
better characterize potential predictability because the
prediction errors at leading times of k steps depend on
the error growth not only from k 2 1 to k, but also from
the initial time to k 2 1.
The averaged G6m over 15 BVs is denoted by G6,
namely,
G6
51
15�15
m51Gm
6 . (10)
FIG. 2. Scatterplots of l of two adjacent breeding cycles for all of the perturbations during the entire period from 1881 to 2000. The lines
in (b)–(f) indicate lmk�1 5 lm
k . Note that the x axis is different between (a) and the other panels. The initial perturbation grows very quickly
at the first rescaling step (Fig. 2a) due to the fast growth of nonlinear instabilities; however, the growth rate varies little after four rescaling
steps, indicating a gradual saturation of the growth of nonlinear instabilities.
302 J O U R N A L O F C L I M A T E VOLUME 24
The term G6 is referred to as the BV rate, the average
growth rate of bred vectors associated with a given ini-
tial state. In total, there are 480 G6 here during the pe-
riod from 1881 to 2000.
3. ENSO BVs and BV rate
In this section, preliminary features of ENSO bred
modes and their variations with initial state, season, and
ENSO signal are explored. The required number of
rescaling steps depends on model dynamics, initial state,
initial perturbation, and other factors. In sensitivity ex-
periments, we found that the fast unstable growth was
saturated after four to six rescaling steps in the HCM.
The scatter-space plot of l of two subsequent rescaling
steps for all of the BVs for the entire period from 1881 to
2000 is shown in Fig. 2. As can be seen in this figure, the
initial perturbation grows very quickly at the first re-
scaling step (Fig. 2a) because of the fast growth of
nonlinear instabilities. However the growth rate varies
little after four rescaling steps, indicating a gradual sat-
uration of the growth of nonlinear instabilities. Thus, we
focused on the BVs of the sixth rescaling step in this
study. Unless otherwise indicated, all of the results of
BVs presented in this paper are from the sixth rescaling
step with the exception of the BV rate that is the cu-
mulative growth rate from rescaling steps 1–6 as defined
above.
a. ENSO bred modes
Figure 3 shows the first EOF modes of BVs for SST
and HC obtained using all of the BVs from 1881 to 2000
(i.e., 480 3 15 BVs), accounting for 61% and 70% of the
total variance, respectively. The first modes, denoted by
EOFtotal, describe a classic ENSO-like structure. Figure
3a for HC has a dipole zonal structure suggesting a
western Pacific ‘‘Rossby wave–like’’ response of one
sign and an eastern Pacific ‘‘Kelvin wave–like’’ response
of the opposite sign (e.g., Tang 2002). Figure 3b for SST
has a large-amplitude signal of one sign located pri-
marily in the equatorial central and eastern Pacific.
These patterns agree with the idea of a heat content
buildup prior to an El Nino as postulated by Jin (1997),
and the delayed oscillator mechanism (Battisti 1988;
Suarez and Schopf 1988); the patterns are also similar
to the BVs from the fully coupled NCEP and the Na-
tional Aeronautics and Space Administration (NASA)
Seasonal-to-Interannual Prediction Project (NSIPP)
GCMs reported in Y06.
When compared with leading singular vectors in many
ENSO models, the patterns in Fig. 3 resemble their final
patterns (e.g., Chen et al. 1997; Xue et al. 1997a; Zhou
et al. 2007; Cheng et al. 2010, hereafter C10). Similarities
between BVs and the final patterns of SVs were also
noticed in CKT03 and Y06.
It has been found in previous work that the SVs are
insensitive to initial conditions in many models (e.g.,
Chen et al. 1997; Xue et al. 1997a; Zhou et al. 2007; C10).
To explore the sensitivity of BV to initial conditions, we
calculated the spatial correlation Rspace between EOFi,
the first EOF mode obtained using 15 BVs for the ith
initial condition (i 5 1, 2, . . . , 480), and EOFtotal, shown
by Fig. 3. If the EOFtotal and the EOFi are denoted by
the normalized one-dimensional vectors eoftotal and eofi,
respectively, the spatial correlation is calculated as fol-
lows:
Rispace 5
1
NP� 1�NP
p51eof
total( p)eof
i( p), (11)
where NP is the number of total model grids over the
model domain, and i 5 1, 2, . . . , 480.
FIG. 3. The first EOF modes of (a) HC and (b) SST. The EOF analysis was employed using all bred vectors from
1881 to 2000. The first modes explain 70% and 61% of the total variance for HC and SST, respectively. The units
are 8C.
1 JANUARY 2011 T A N G A N D D E N G 303
The spatial correlation measures the similarity be-
tween the dominant direction in the space spanned by
BVs at different time instances and the dominant di-
rection in the space spanned by BVs computed over
a long time period. The result is shown in Fig. 4. For
most cases (over 90%), the absolute value of the spatial
correlation coefficient is over 0.90, with an overall av-
erage value above 0.85 for all cases for both HC and
SST. As argued by CKT03, the positive phase of the
same pattern should be regarded as equivalent to the
negative phase, at least from a linear system perspective.
In addition, the pattern sign can be arbitrary from the
view of EOF analysis. We also performed bootstrap
experiments [i.e., randomly choose two initial condi-
tions (among 480)], and the spatial correlation of the
leading EOF mode was calculated between the two ini-
tial conditions [i.e., eofj instead of eoftotal in (11)]. The
process was repeated 1000 times. The results showed
that 70% and 93% of the absolute values of the correlation
coefficient were over 0.652 for SST and HC, respectively.
Thus, the preliminary features of BVs, as represented by
the first EOF mode (equivalent to the average), are not
highly sensitive to initial conditions. This conclusion is
further confirmed by Fig. 5, the pattern of the first EOF
modes obtained using all of the BVs in four different
calendar months, indicating the seasonal independence of
the BVs.
The above finding on the insensitivity of BVs to ini-
tial conditions results from the similarity of the EOF-
characterized dominant structure in the space spanned
by the BVs of different initial conditions. Further exami-
nation was conducted on individual BVs at different initial
conditions. In other words, is an individual BV at the
FIG. 4. The spatial correlation between EOFtotal and EOFi, where EOFtotal is the first EOF mode obtained using all
BV vectors (480 3 15), whereas EOFi is the first EOF mode obtained using 15 BVs for the initial condition i (i 5 1,
2, . . . , 480).
2 The value of 0.65 was chosen arbitrarily, a very large value
compared with the threshold value of statistical significance at the
confidence level of 99%.
304 J O U R N A L O F C L I M A T E VOLUME 24
initial time t1 similar to another BV at its initial time t2? To
explore this question, we randomly chose two individual
BVs at two different initial conditions and calculated their
spatial correlation coefficient. This process was repeated
1000 times, as shown in Fig. 6. The correlation coefficients
are almost evenly distributed across the range from the
maximum (1) to the minimum (21), indicating that an
individual BV is dependent on its initial conditions.
The average of the ensemble BVs had a stable struc-
ture independent of the initial state whereas the in-
dividual member of ensemble BVs was sensitive to the
initial state. One physical interpretation for this is that
the mean of ensemble BVs is probably dominated by
time-independent model dynamics (e.g., the delayed-
oscillator mechanism), whereas the individual BV mem-
ber is greatly impacted by an initial random perturbation.
In Toth and Kalnay’s approach, the initial perturbation
came from multiple modes orthogonalizing in various
ways that were not completely random in nature; thus, the
BV derived from their approach is probably equivalent to
the mean of the ensemble of BVs here.
b. Variations of BV rate
The BV rate G6, defined in section 2b, measures the
mean growth rate of all of the BVs for a given initial
state. Shown in Fig. 7 is the average G6 of all of the BVs
for four different calendar months. As can be seen, the
greatest BV rates occur in January and April, during
which the model runs through the boreal spring and
summer (because of the 6 rescaling steps for each BV,
equivalent to 6 months). This is in good agreement with
the variation of the optimal error growth rate in the ZC
model (e.g., CKT03; C10). The seasonal dependence of
error growth might explain why ENSO prediction skill
often drops remarkably when predictions go through the
spring (i.e., the ‘‘spring barrier’’). This increase in error
is due to the fact that the intertropical convergence zone
(ITCZ) is closest to the equator during the spring, sus-
taining unstable conditions and enhancing convective
instability. In contrast, the ocean–atmosphere interac-
tion is strong during the summer due to the relatively
large vertical temperature gradient and ocean upwelling
(e.g., Xue et al. 1997a), which lead to strong baroclinic
instability.
Because of the existence of decadal/interdecadal
variations in ENSO variability and prediction skills (e.g.,
Tang et al. 2008a), it is interesting to examine temporal
variations in the BV rate. Figures 8a,b show variations in
the BV rate and the Nino-3.4 SSTA index from 1881 to
2000, for which a 2-yr running mean has been applied to
FIG. 5. The first EOF mode of HC obtained using all of the BVs of four different calendar months.
1 JANUARY 2011 T A N G A N D D E N G 305
address interannual and longer signals. As previously
indicated, the interannual variability is obvious in both
the BV rate and in the Nino-3.4 SSTA index. Further
scrutiny of Figs. 8a,b reveals that the interannual vari-
ability of the BV rate has decadal/interdecadal varia-
tion. For example, the magnitude of the BV rate is
visibly smaller during the 1880s and 1980s than during
other periods. The variability of the BV rate at various
time scales is more obvious in the wavelet analysis
shown in Fig. 8c. The local wavelet power spectrum
clearly indicates that significant periods were localized
in time and varied from 2 to 16 yr from 1881 to 2000. A
striking feature in Figs. 8c,d is the strong interannual
variability and relatively weak decadal variability. Com-
paring Figs. 8c with 8d reveals similar oscillatory char-
acteristics between the BV rate and ENSO variability,
suggesting an association of the BV rate with the ENSO
background.
The dependence of the growth rate of BVs on the
ENSO phase has been addressed in CKT03 and Y06.
They found that a large error growth tends to occur in
a neutral or onset/breakdown stage of an ENSO event,
such as 3–4 months prior to the peak phase, whereas
a small error growth often occurs in an ENSO peak
phase. In their work, the model integration was used as
the reference trajectory for their BV analysis, as men-
tioned in section 2b. The relationship between the BV
rate and ENSO phase was then explored by comparing
our BVs with the realistic ENSO evolution as the initial
reference trajectory.
Figure 9 shows the average BV rate as a function of
ENSO phase. The ENSO events are binned into nine
categories based on the Nino-3.4 SSTA index. Three
kinds of events, El Nino, La Nina, and neutral events,
are explored here. They are classified by Nino-3.4 SSTA
index (8C) greater than 2.0, 1.5, and 1.0; less than 22.0,
21.5, and 1.0; and in the range of (20.5, 0.5), (20.3, 0.3),
and (20.1, 0.1), respectively. Considering 6 rescaling
steps (equivalent to 6 months) for each BV, we selected
the BV rate of the 3 months prior to the time when the
FIG. 6. Correlation between two arbitrary BVs from two arbitrary initial conditions. These arbitrary BVs and initial
conditions were randomly chosen, and this process was repeated 1000 times.
306 J O U R N A L O F C L I M A T E VOLUME 24
criteria value was met. Thus, the BV rate shown in
Fig. 9 actually represents the error growth of a 6-month
duration (i.e., 3-month prior to and after the occur-
rence of an ENSO event). As can be seen in Fig. 9,
there was a small error growth rate during the peak
ENSO stages (El Nino and La Nina), while a large
error growth rate occurred during the neutral stages.
The stronger the background SSTA of ENSO, the
smaller the BV rate. These results are generally con-
sistent with those in CKT03 and Y06 as well as former
SV studies (e.g., Chen et al. 1997; Zhou et al. 2007).
Both CKT03 and Zhou et al. 2007 analytically derived
this converse relationship between the error growth
rate and the ENSO signal by using a delay-oscillator
model (Suarez and Schopf 1988). This relationship
suggests that a strong ENSO signal tends to reside in
a dynamical stable regime, whereas a weak signal is
probably more associated with an instable dynamical
process. Another possibility is that the model ENSO
strength is bounded above so that error growth can take
place only in one direction (e.g., neutral phase) when
one is near the extreme, resulting in a situation where
instability might remain large in the direction toward
neutral. In addition, Fig. 9 shows that the BV rate is
smaller during El Nino than during La Nina, possibly
because the tropical wave instability is more active
when the cold tongue is well established in the east
(Y06).
The converse relationship between the BV rate and
ENSO signal is further demonstrated in Fig. 10, the
scatterplot of the BV rate against the ENSO signal. The
ENSO signal here is defined by the variance of the ob-
served Nino-3.4 SSTA index, computed using a 20-yr
running window from 1881 to 2000.3 A strong converse
relationship is observed in Fig. 10 with a correlation
coefficient of 20.71 between the ENSO signal and the
BV rate. A similar result can be obtained using other
measures of the ENSO signal, such as the total spectrum
power of Nino-3.4 SSTA index at the frequencies of
2–5 yr or the variance explained by the first EOF mode
of SSTA (HCA).
FIG. 7. The average BV rate for all of the initial conditions for
January, April, July, and October.
FIG. 8. Variation in (a) the BV rate and (b) the Nino-3.4 SSTA
index from 1881 to 2000. The wavelet power spectrum of (c) the BV
rate and (d) the Nino-3.4 SSTA using the Morlet wavelet. The thick
contour encloses regions of greater than 95% confidence using
a red-noise background spectrum. The solid smooth curves in the
left and right corners indicate that edge effects become important.
The time window used here was 2ffiffiffi2p
T, a function of period T.
3 The variance obtained during a 20-yr window is plotted at the
middle point of the window. For example, the variance at 1890 is
calculated using the samples from 1881 to 1900. The 20-yr window
is shifted by 3-months each time starting from 1881 to 2000. The BV
rate plotted is the average value of BVs of each window.
1 JANUARY 2011 T A N G A N D D E N G 307
4. ENSO predictability
In this section, we will explore ENSO predictability.
Emphasis was placed on identifying a measure calcu-
lated from BVs that was able to quantify potential pre-
dictability and had a relationship with actual prediction
skill. This relationship offers a practical means of esti-
mating the confidence level of an ENSO forecast using
the HCM.
The retrospective ENSO prediction was performed for
the period from 1881 to 2000 using the HCM. A total of
480 forecasts, initialized from January 1881 to October
2000, were run starting at 3-month intervals (1 January,
1 April, 1 July, and 1 October), and continued for 12 months
for the HCM. The SST assimilation was performed to
initialize the forecasts, as introduced in section 2a.
a. Bred vector and BV rate
Figure 11 shows the average correlation R and RMSE
of SSTA predictions against the observed values over
lead times of 1–6 months [i.e., (1/6)�6
l51Rl(RMSE
l),
where l is the index of lead month] for the period from
1881 to 2000. As indicated, the best correlations occur in
the equatorial central and eastern Pacific (as in many
ENSO models; e.g., Goddard et al. 2001; Tang et al.
2008a), where there was the largest error growth, as
shown in Fig. 3. This seemingly contradicts the wide-
spread perception that good prediction ability should
correspond with a small error growth rate (e.g., Farrell
1990; Tang et al. 2008b; Orrell 2003). One probable
reason is that error growth, as shown in Fig. 3, can be
decomposed into linear growth and nonlinear growth.
The linear errors only impact amplitude prediction,
whereas nonlinear errors impact phase prediction. Thus,
the correlation is still likely to be high if the large error
growth rate is due mainly to linear instabilities. Tang and
Deng (2010) found that, indeed, the region of high cor-
relation coefficients has relatively weak nonlinearity. This
argument is also supported by the fact that the region of
good correlations often has a large RMSE in ENSO
models, as shown in Fig. 11b.
Figure 12 shows the scatterplot of the correlation and
the RMSE of individual predictions against the corre-
sponding BV rates using the same initial conditions ap-
plied to both the BVs and the predictions. The RMSE/
correlation value of individual predictions (RMSEIP/
CIP) was used to measure the error (correlation) of an
individual forecast for all leads L (512 months), defined as
RMSEIP 5
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
L� 1�t5L
t51[Tp(t)� To(t)]2
vuutand (12)
CIP 5
�L
t51(T
pt � mp)(To
t � mo)ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�L
t51(T
pt � mp)2
vuutffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�L
i51(To
t � mo)2
vuut. (13)
Here, T is the index of Nino-3.4 SSTA, p denotes
forecasts, o denotes observations, t is the lead time of
the forecast, mp is the mean of the forecasts, mo is the
FIG. 9. The average BV rate as a function of the ENSO phase,
indicating the error growth rate during La Nina, neutral, and
El Nino events. Three kinds of El Nino (La Nina) cases were ex-
plored with Nino-3.4 SSTA index greater (less) than 1.0 (21.0), 1.5
(21.5), and 2.0 (22.0), respectively; whereas three kinds of neutral
cases were explored using the ranges of Nino-3.4 SSTA index be-
tween (20.5, 0.5), (20.3, 0.3), and (20.1, 0.1), respectively.
FIG. 10. Scatterplot of the BV rate against the ENSO signal.
308 J O U R N A L O F C L I M A T E VOLUME 24
mean of observations, and L is the number of samples
used.
Figure 12 includes all of the initial conditions from
1881 to 2000, showing little visible relationship between
the BV rate and prediction skill and suggesting that the
BV rate probably is not a good measure of potential
predictability for this HCM. However, one should take
caution when interpreting this figure because of the
small sample size used in making this calculation, es-
pecially for the correlation coefficient. We next explore
the other measure of potential predictability and ex-
amine its relationship to prediction skills.
b. Relative entropy derived from BVs
Recently, a new theoretical framework for measuring
the uncertainty of predictions has been developed and
applied to examine atmospheric and climate potential
predictabilities (Schneider and Griffies 1999; Kleeman
2002, 2008; Tippett et al. 2004; Tang et al. 2005, 2008b;
DelSole 2004, 2005). The approach is built on informa-
tion theory (Cover and Thomas 1991). It has been argued
that the relative entropy (RE) of an individual prediction,
defined as the difference between the climatological
probability density function (PDF) and the prediction
PDF, is a good measure of its prediction utility. A larger
RE indicates that more useful information is being sup-
plied by the prediction, which could be interpreted as
making it more reliable.
In the case where the PDFs are Gaussian, which is
approximately the case in many practical instances (in-
cluding ENSO prediction), the relative entropy RE of
an individual prediction may be expressed exactly in
terms of the prediction variance sp2, evaluated typically
by ensemble members, the model climatological vari-
ance sq2, obtained from a model climatological run, and
the difference mp 2 mq of the ensemble and the clima-
tological means (Kleeman 2002):
RE 51
2log
s2q
s2p
1s2
p
s2q
1(mp � mq)2
s2q
� 1
" #. (14)
Figure 13 is as in Fig. 12, but using the RE of Nino-3.4
SSTA index instead of the BV rate. Here the climato-
logical variance and mean were obtained from samples
of all BVs for all initial conditions from 1881 to 2000
(i.e., 480 3 15 samples), as in Tippett et al. (2004) and
Tang et al. (2005). As can be seen in Fig. 13, a large RE
often led to a good prediction skill (i.e., a small RMSE
and a large correlation); however, when the RE was
small, the prediction skill seemed much more variable.
This result is very similar to the so-called triangular re-
lationship that has been found to characterize the re-
lationship between ensemble spread and prediction
skill in ensemble weather forecast. This finding also re-
sembled the relationship between ensemble signal and
prediction skill in ensemble climate prediction (ECP;
e.g., Buizza and Palmer 1998; Xue et al. 1997b; Moore
and Kleeman 1998; Tang et al. 2008b). For instance, it
was found in ECP that, when the ensemble signal is
large, the prediction is invariably good; whereas when it
is small, the prediction can be much more variable. Thus,
we also used the ‘‘triangular relationship’’ to describe
the relationship between the RE and prediction accu-
racy in this study. Figure 13 suggests that the RE derived
from BVs could be an effective indicator of ENSO pre-
diction skill.
The importance of RE in measuring potential pre-
dictability has been addressed in previous studies (e.g.,
Kleeman 2002, 2008; Tippett et al. 2004; Tang et al. 2005,
2008b), but all of these used prediction ensemble to
FIG. 11. (a) Correlations and (b) RMSEs of predicted Nino-3.4
SSTA against the observed values for the past 120 yr from 1881 to
2000, averaged over lead times of 1–6 months.
1 JANUARY 2011 T A N G A N D D E N G 309
calculate RE. Here we used BVs to calculate RE and
obtained similar results. This is probably due to the in-
trinsic coherence of BVs with ensemble predictions.
From the point of calculation expense, the BV-based
RE is comparable to the ensemble-prediction-based RE
when the BV analysis is incorporated with the ensemble
run. However, if the ensemble prediction system is in-
dependent to the generation of the optimal perturbation
modes (e.g., singular vector), the BV-based RE is much
cheaper than the ensemble-prediction-based RE.
c. Decadal/interdecadal variation in relative entropyand predictability
In many ENSO models, ENSO predictability clearly
has a decadal/interdecadal variation (e.g., Chen et al.
2004; Tang et al. 2008a). It was found that the inter-
decadal variation in ENSO predictability was in good
agreement with the signal of interannual variability
and with the degree of asymmetry of an ENSO system
(e.g., Tang et al. 2008a; Tang and Deng 2010). Below
the relationship between relative entropy and ENSO
predictability at decadal/interdecadal time scales is
examined.
Figure 14 shows the average R and RMSE values for
the Nino-3.4 SSTA index with a 1–12-month lead,
measured by a 20-yr running window that moved at 1-yr
intervals from 1881 to 2000 (i.e., 1881–1900, 1882–1901,
1883–1902, . . . , 1980–99, 1981–2000), where the predic-
ted Nino-3.4 SSTA is compared against the observed
value.4 The average relative entropy over the running
window is also plotted in Fig. 14 (solid line). As can be
seen in this figure, the prediction skills show a pro-
nounced interdecadal variation. Generally there was a
high prediction skill in the late nineteenth century and in
the middle to late twentieth century, and a low skill from
1901 to 1960. Such an interdecadal variation feature was
also manifested in the RE. The correlations between the
RE and the correlation coefficient and between the RE
and the RMSE were 0.76 and 20.66, respectively, thus,
the larger the RE, the larger the correlation and the
smaller the RMSE.
Further examination of RE might provide some in-
sights into interdecadal variations in ENSO predict-
ability. The first two terms on the right-hand side (rhs) of
(14) are determined by the climatological variance and
predictive variance. Because variance of climatology is
time invariant, these terms represent a measurement of
FIG. 12. Scatterplot of (a) correlations and (b) RMSEs of each individual prediction against the BV rate for the
corresponding initial condition.
4 The average skills, as a function of the index of window, were
obtained by (1/12)�12
l 5 1RLl (RMSEL
l ), where l is the index of the
lead month and L is an index of a 20-yr window that shifts by 1 yr
from the beginning to the end during the period from 1881 to 2000
(January 1881, April 1881, . . . , October 2000). The term RlL(RMSEl
L)
is the skill for the window L measured using 20-yr samples, as a
function of the lead time l.
310 J O U R N A L O F C L I M A T E VOLUME 24
the dispersion or spread of the BVs. The third term on
the rhs of (14) is governed by the amplitude of the mean
of the BVs. We refer to the first two terms minus 1 as the
dispersion component (DC) and the third term as the
signal component (SC).
Figure 15 shows the variations in DC and SC against
RE for the period from 1881 to 2000. This figure reveals
that DC is significantly larger than SC, and RE is dom-
inated by DC. As can be seen in Fig. 15a, RE and DC
vary linearly with a slope of unity. The correlation co-
efficient between R and DC is 0.97. In contrast to the
good relationship between DC and RE, the relation
between SC and RE was much less significant. Figure 13
indicates that the triangular relationship between the
RE and prediction skill is equivalent to the triangular
relationship between ensemble spread and skill that was
found in some numerical weather models.
It should be noticed that Fig. 15 differs from an earlier
result reported by Tang et al. (2005). They found that
the SC plays a dominant role in the RE s for two ENSO
models. This is most likely due to disparities with models
and methods used in these two cases. In this study, we
used a different HCM. In particular, BVs were used to
calculate the RE rather than the ensemble prediction
itself used in Tang et al. (2005). Figures 15 and 14 in-
dicate a possible mechanism of interdecadal variation in
ENSO predictability; namely, when the ENSO cycle
resides in a phase that has a relatively large DC, it is
more predictable and vice versa.
5. Discussion and summary
The breeding method has been widely applied in a
variety of fields including atmospheric and climate pre-
dictability, data assimilation, and ensemble prediction.
FIG. 13. Scatterplot of (a) correlations and (b) RMSEs of each individual prediction against the relative entropy for
the corresponding initial condition.
FIG. 14. Variations in the relative entropy (solid), the average
correlation coefficient (dashed), and RMSE (dotted) over 1–
12-month lead times from 1881 to 1990. Normalization was done
for each variable prior to plotting.
1 JANUARY 2011 T A N G A N D D E N G 311
In this study, we applied the breeding method to a hy-
brid coupled model and demonstrated the feasibility of
using BVs to facilitate ENSO predictability studies.
Specifically we performed, for the first time, each BV
starting from a realistic ENSO background for the past
120 yr from 1881 to 2000. Meanwhile, retrospective
forecasts of the past 120 yr were also conducted. The
absence of stochastic atmospheric transients in this hy-
brid coupled model facilitated the ability of the breeding
method to identify slowly growing error modes. The
results show that the breeding method was able to detect
climatically relevant coupled instabilities present in this
coupled model.
The characteristics of ENSO BVs and BV rates were
analyzed. It was found that the preliminary BV mode
had an ENSO-like structure with a dipole zonal struc-
ture between the western Pacific and the eastern Pacific
in the subsurface heat content field and a large-amplitude
signal in the equatorial central and eastern Pacific. This
structure is similar to the final pattern responding to the
first singular vector and also reminiscent of the physical
processes controlled by the delayed-oscillator mecha-
nism. The dominant structure is insensitive to initial
conditions like singular vectors in many ENSO models.
However, the spatial structure of each individual BV is
sensitive to the initial conditions.
The seasonal dependence of the error growth rate
(BV rate) was investigated. The largest BV rates oc-
curred when the model runs through the boreal spring
and summer (i.e., spring barrier). The BV rate was also
strongly dependent on the ENSO phase. A small error
growth rate occurred during the peak ENSO stages,
whereas a large error growth rate occurred during the
neutral stages. A significant converse relationship exis-
ted between the signal of the background SSTA and the
error growth rate: the stronger the background SSTA of
ENSO, the smaller the BV rate, and vice versa. This
converse relationship is primarily due to the nonlinear
relationship (e.g., exponential) between the signal and
growth rate, as indicated by theoretical analyses (CKT03;
Zhou et al. 2007).
To explore ENSO predictability, we conducted ret-
rospective ENSO predictions for the period from 1881
to 2000. The BV rate was analyzed in terms of its ability
to measure the model prediction skill. It was found that
an overall relationship between the BV rate and pre-
diction skill was absent in this model. Furthermore, we
calculated and analyzed the relative entropy from BVs.
The relative entropy was found to be an effective in-
dicator of prediction utility. The relationship between
the relative entropy and prediction skill varies at a vari-
ety of time scales. For individual cases, it appeared to be
like a ‘‘triangular relationship’’: large relative entropy
often led to a good prediction skill; however, when the
relative entropy was small, the prediction skill seemed
much more variable. At decadal/interdecadal scales, there
was consistent variation in prediction skill and in relative
entropy. In the late-nineteenth and the late-twentieth
FIG. 15. The relationship of the (a) dispersion component and the (b) signal component to RE.
312 J O U R N A L O F C L I M A T E VOLUME 24
centuries, the relative entropy was large and the model
had a high correlation and low RMSE. During other
periods, the relative entropy was small, and the model
showed the poor prediction skill. Furthermore, it was
found that relative entropy was dominated by the spread
of BVs, whereas the mean of BVs had little contribution
to RE.
The coupled model used in this study was a hybrid
coupled in absence of the fastest-growth weather in-
stabilities inherent to stochastic atmospheric transients.
While such a hybrid model facilitates the detection of
ENSO error modes by using the breeding method, it
lacks some necessary physical and dynamical processes
that exist in the real world. It would be interesting to
validate the results presented in this study with com-
plicated GCM models, a line of research that we will
pursue in the future. In addition, the window length of
20 yr used in this study is slightly arbitrary and was
motivated by previous work (e.g., Chen et al. 2004; Tang
et al. 2008a). Nevertheless, this study analyzed ENSO
BVs using realistic ENSO backgrounds as reference
trajectories, and, in particular, explored the relationship
between BVs and actual prediction skill for a long pe-
riod of over 100 yr. The results offer valuable insight to
ENSO predictability and have practical significance to
ensemble prediction.
Acknowledgments. This work is supported by the Ca-
nadian Foundation for Climate and Atmospheric Sci-
ences (CFCAS) through Grant GR-7027 and the Canada
Research Chair program. We thank the anonymous re-
viewers for their constructive comments, which helped to
improve this manuscript.
REFERENCES
Battisti, D. S., 1988: Dynamics and thermodynamics of a warming
event in a coupled tropical atmosphere-ocean model. J. At-
mos. Sci., 45, 2889–2919.
Buizza, R., and T. N. Palmer, 1998: Impact of ensemble size on
ensemble prediction. Mon. Wea. Rev., 126, 2503–2518.
Cai, M., E. Kalnay, and Z. Toth, 2003: Bred vectors of the Zebiak–
Cane model and their application to ENSO predictions.
J. Climate, 16, 40–56.
Chen, D., M. A. Cane, A. Kaplan, S. E. Zebiak, and D. Huang,
2004: Predictability of El Nino in the past 148 years. Nature,
428, 733–736.
Chen, Y.-Q., D. S. Battisti, T. N. Palmer, J. Barsugli, and
E. S. Sarachik, 1997: A study of the predictability of tropical
Pacific SST in a coupled atmosphere–ocean model using sin-
gular vector analysis: The role of the annual cycle and the
ENSO cycle. Mon. Wea. Rev., 125, 831–845.
Cheng, Y., Y. Tang, X. Zhou, P. Jackson, and D. Chen, 2010:
Further analysis of singular vector and ENSO predictability in
the Lamont model—Part I: Singular vector and the control
factors. Climate Dyn., 35, 807–826, doi:10.1007/s00382-009-
0595-7.
Corazza, M., E. Kalnay, D. J. Patil, S. C. Yang, and R. Morss, 2003: Use
of the breeding technique to estimate the structure of the analysis
‘‘error of the day.’’ Nonlinear Processes Geophys., 10, 1–11.
Cover, T. M., and J. A. Thomas, 1991: Elements of Information
Theory. Wiley, 576 pp.
DelSole, T., 2004: Predictability and information theory. Part I:
Measures of predictability. J. Atmos. Sci., 61, 2425–2440.
——, 2005: Predictability and information theory. Part II: Imper-
fect forecasts. J. Atmos. Sci., 62, 3368–3381.
Deng, Z., and Y. Tang, 2009: Reconstructing the past wind stresses
over the tropical Pacific Ocean from 1875 to 1947. J. Appl.
Meteor. Climatol., 48, 1181–1198.
——, ——, and X. Zhou, 2008: The retrospective prediction of
El Nino–Southern Oscillation from 1881 to 2000 by a hybrid
coupled model: (I) Sea surface temperature assimilation with
ensemble Kalman filter. Climate Dyn., 32, 397–413, doi:10.1007/
s00382-008-0399-1.
Evensen, G., 2003: The ensemble Kalman filter: Theoretical
formulation and practical implementation. Ocean Dyn., 53,
343–367.
Fan, Y., M. R. Allen, D. L. T. Anderson, and M. A. Balmaseda,
2000: How predictability depends on the nature of uncertainty
in initial conditions in a coupled model of ENSO. J. Climate,
13, 3298–3313.
Farrell, B., 1990: Small error dynamics and the predictability of
atmospheric flows. J. Atmos. Sci., 47, 2409–2416.
Goddard, L., S. J. Mason, S. E. Zebiak, C. F. Ropelewski, R. Basher,
and M. A. Cane, 2001: Current approaches to seasonal-to-
interannual climate predictions. Int. J. Climatol., 21 (9), 1111–
1152.
Jin, F.-F., 1997: An equatorial ocean recharge paradigm for ENSO.
Part I: Conceptual model. J. Atmos. Sci., 54, 811–829.
Kalnay, E., 2003: Atmospheric Modeling, Data Assimilation, and
Predictability. Cambridge University Press, 341 pp.
Kleeman, R., 2002: Measuring dynamical prediction utility using
relative entropy. J. Atmos. Sci., 59, 2057–2072.
——, 2008: Limits, variability, and general behavior of statistical
predictability of the midlatitude atmosphere. J. Atmos. Sci.,
65, 263–275.
Lorenz, E. N., 1965: A study of the predictability of a 28-variable
atmospheric model. Tellus, 17, 321–333.
Moore, A. M., and R. Kleeman, 1996: The dynamics of error
growth and predictability in a coupled model of ENSO. Quart.
J. Roy. Meteor. Soc., 122, 1405–1446.
——, and ——, 1997a: The singular vectors of a coupled ocean-
atmosphere model of ENSO. Part I: Thermodynamics, ener-
getics and error growth. Quart. J. Roy. Meteor. Soc., 123, 953–981.
——, and ——, 1997b: The singular vectors of a coupled ocean-
atmosphere model of ENSO. Part II: Sensitivity studies and
dynamical interpretation. Quart. J. Roy. Meteor. Soc., 123,
983–1006.
——, and ——, 1998: Skill assessment for ENSO using ensemble
prediction. Quart. J. Roy. Meteor. Soc., 124, 557–584.
Orrell, D., 2003: Model error and predictability over different
timescales in the Lorenz ’96 systems. J. Atmos. Sci., 60, 2219–
2228.
Pena, M., and E. Kalnay, 2004: Separating fast and slow modes in
coupled chaotic systems. Nonlinear Processes Geophys., 11,
319–327.
Schneider, T., and S. M. Griffies, 1999: A conceptual framework for
predictability studies. J. Climate, 12, 3133–3155.
1 JANUARY 2011 T A N G A N D D E N G 313
Smith, T. M., and R. W. Reynolds, 2004: Improved extended re-
construction of SST (1854–1997). J. Climate, 17, 2466–2477.
Suarez, M. J., and P. S. Schopf, 1988: A delayed action oscillator for
ENSO. J. Atmos. Sci., 45, 3283–3287.
Tang, Y., 2002: Hybrid coupled models of the tropical Pacific—
Interannual variability. Climate Dyn., 19, 331–342.
——, and Z. Deng, 2010: Low-dimensional nonlinearity of ENSO
and its impact on predictability. Physica D, 239, 258–268.
——, R. Kleeman, and A. M. Moore, 2005: On the reliability of
ENSO dynamical predictions. J. Atmos. Sci., 62, 1770–1791.
——, ——, and S. Miller, 2006: ENSO predictability of a fully
coupled GCM model using singular vector analysis. J. Climate,
19, 3361–3377.
——, Z. Deng, X. Zhou, Y. Cheng, and D. Chen, 2008a: Inter-
decadal variation of ENSO predictability in multiple models.
J. Climate, 21, 4811–4833.
——, H. Lin, and A. Moore, 2008b: Measuring the potential pre-
dictability of ensemble climate predictions. J. Geophys. Res.,
113, D04108, doi:10.1029/2007JD008804.
Thompson, C. J., 1998: Initial conditions for optimal growth in
a coupled ocean-atmosphere model of ENSO. J. Atmos. Sci., 55,
537–557.
Tippett, M. K., R. Kleeman, and Y. Tang, 2004: Measuring the
potential utility of seasonal climate predictions. Geophys. Res.
Lett., 31, L22201, doi:10.1029/2004GL021575.
Toth, Z., and E. Kalnay, 1993: Ensemble forecasting at NMC—The
generation of perturbations. Bull. Amer. Meteor. Soc., 74,
2317–2330.
——, and ——, 1997: Ensemble forecasting at NCEP and the
breeding method. Mon. Wea. Rev., 125, 3297–3319.
Xue, Y., M. A. Cane, and S. E. Zebiak, 1997a: Predictability of
a coupled model of ENSO using singular vector analysis. Part
I: Optimal growth in seasonal background and ENSO cycle.
Mon. Wea. Rev., 125, 2043–2056.
——, ——, ——, and T. N. Palmer, 1997b: Predictability of a cou-
pled model of ENSO using singular vector analysis. Part II:
Optimal growth and forecast skill. Mon. Wea. Rev., 125, 2057–
2073.
Yang, S.-C., M. Cai, E. Kalnay, M. Rienecker, G. Yuan, and Z. Toth,
2006: ENSO bred vectors in coupled ocean–atmosphere general
circulation models. J. Climate, 19, 1422–1436.
Young, R. M. B., and P. L. Read, 2008: Breeding and predictability
in the baroclinic rotating annulus using a perfect model.
Nonlinear Processes Geophys., 15, 469–487.
Zebiak, S. E., and M. A. Cane, 1987: A model El Nino–Southern
Oscillation. Mon. Wea. Rev., 115, 2262–2278.
Zhou, X., Y. Tang, and Z. Deng, 2007: The impact of atmospheric
nonlinearities on the fastest growth of ENSO prediction
error. Climate Dyn., 30, 519–531, doi:10.1007/s00382-007-
0302-5.
314 J O U R N A L O F C L I M A T E VOLUME 24